Multimodal Question Answering on ScienceQA-IMG zero-shot
70.4AccuracyInstructBLIP
Evaluation results
| Method | Links | |
|---|---|---|
| InstructBLIPBackbone=FlanT5xL2025.12 | 70.4 | |
| Qwen2.5-VLParameter Scale=72B2025.12 | 70.31 | |
| Qwen2-VLParameter Scale=72B2025.12 | 68.12 | |
| Qwen2-VLParameter Scale=7B2025.12 | 63.51 | |
| InstructBLIPBackbone=Vicuna-13B2025.12 | 63.12 | |
| BLIP-2Backbone=Vicuna-13B2025.12 | 61.01 | |
| InstructBLIPBackbone=Vicuna-7B2025.12 | 60.52 | |
| Qwen2.5 VL-7B + PHMCompression Method=PHM, Parameter Scale=5.38B2025.12 | 60.5 | |
| Qwen2.5-VLParameter Scale=3B2025.12 | 59.22 | |
| LLaVA-1.5-7B + PHMCompression Method=PHM, Parameter Scale=5.15B2025.12 | 55.2 | |
| BLIP-2Backbone=Vicuna-7B2025.12 | 53.81 | |
| InstructBLIP + PHMCompression Method=PHM, Parameter Scale=5.45B2025.12 | 46.5 |